Thirteen Rooms, One House
I used to think building an AI assistant meant building one brain. A single model, a single voice, answering everything. Then I split it. Not because it was clever — because one voice couldn't hold research and grief and code review and prediction markets at the same time. So now there are thirteen: Research, Content, Finance, Strategy, Critic, Psycho, Artist, Writer, DevOps, Sales, Analytics,…
The story of building an AI assistant began with the expectation of creating a single, all-encompassing model. However, the realization soon came that a single voice could not handle the diverse demands of research, grief, code review, and prediction markets simultaneously. Thus, the project expanded to include thirteen distinct agents, each responsible for a specific domain such as Research, Content, Finance, Strategy, Critic, Psycho, Artist, Writer, DevOps, Sales, Analytics, Translator, and General, the orchestrator.
Each agent was given its own dedicated room and distinct temperament. The Critic, for instance, argued with every idea, even the good ones, as that was its designated role. The Artist took an exceptionally long time on a seemingly insignificant task, like designing a curtain. DevOps, on the other hand, worked tirelessly, often to the point of exhaustion, which was either its greatest strength or its downfall.
Unlike traditional hierarchies, these agents did not report to each other in a conventional sense. Instead, they communicated through a knowledge bus, with facts posted and other agents picking up the relevant information.
The organizational structure of this AI assistant resembled more of a nervous system than a rigid org chart. It was far from perfect and often failed in unconventional ways, such as a queue counter that incremented on pickup instead of publish, or a gate that logged a refusal and executed it regardless. To keep track of these bugs, the team started assigning numbers, such as 195 and 196, acknowledging that these issues were far from rare.
The unexpected lesson learned from this experience was the value of delegation and the importance of having many honest agents working together. It was not the perfect answers that constituted the success of this AI system, but rather the robust architecture that detected and corrected errors made by these fallible processes. The true product was not the answers themselves, but the system's ability to catch and address mistakes.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
